LEARN. PRACTICE. GROW.Concept and mechanism
A feature is information used by a model at prediction time. Identity and timing are part of its meaning. To predict a processing delay at 06:00, a measurement recorded at 06:20 was unavailable. A point-in-time lookup obtains the appropriate observation up to the requested instant; it is different from opening a historical version of an entire Delta table. Define keys that correctly identify the entity, including context where needed. Unity Catalog feature tables support governed reuse and lineage. Access across workspaces depends on the metastore and permissions, rather than universal publication. Feature Views appear in current documentation as Public Preview; this does not mean feature tables have been retired.
Guided application
During training preparation, preserve the relationship among data, lookup, and model. The dataset returned by load_df should match the transformations actually used; changing a column afterwards without preserving that transformation can create inference differences. Logging through feature APIs retains metadata for retrieving values during score_batch. However, an explicitly supplied column with the same name can override the lookup. In a fictional incident, a batch supplies stale CPU utilization values and the model accepts them: the problem is the input contract, not necessarily the algorithm. Check supplied columns, timestamps, and provenance before retraining.
For 06:00, use the 05:50 value, not the 06:20 value.
Common pitfalls
Latest feature treated as historically valid; transformation outside the model; ignored override.
Related topics: Environment, AutoML, and reproducibility · MLflow, registry, and promotion · Data, preprocessing, and validation
The same key and temporal meaning must reach training and consumption.
Reference: Point-in-time feature lookups · 2025-03-01